feat(evaluation-system): add claude -p eval engine and response cache

[cmschuetz]

Introduce the foundational eval-engine layer that runs a single-shot model
invocation through the Claude Code CLI in print mode (claude -p) rather than
the Anthropic SDK, so the harness's real output schemas are exercised end to
end. The engine forces schemas with --json-schema, reads cost/latency/usage
from the --output-format json envelope, content-addresses responses in a
file-backed cache, and stays fully hermetic under bun test by faking the
subprocess boundary. This is the substrate every later deliverable builds on.

Summary

New src/eval/ module: engine.ts (buildArgs/parseEnvelope/runClaude/
isAvailable), types.ts (EvalRequest/EvalResult/Envelope/Spawn), and
cache.ts (sha256 content-addressed response cache). Envelope parsing prefers
structured_output, falls back to JSON.parse(result), grades error/mismatch/
unparseable cases as ok:false without throwing, and surfaces an absent CLI as
a distinct skip signal. All tests inject a fake Spawn — no real claude, no
new runtime dependency.

Acceptance criteria

• runClaude invokes only claude; argv includes --print --output-format json --json-schema … --model … --strict-mcp-config --settings … plus system-prompt flags when provided.
• Success envelope with structured_output yields EvalResult{ok:true,…}; missing structured_output falls back to JSON.parse(result).
• is_error, non-success subtype, unparseable stdout, and schema-nonconforming output each yield ok:false with a descriptive error and no throw.
• Response cache keys on prompt+systemPrompt+model+schema+tools+settings: identical request hits, any keyed-field change misses.
• Absent claude CLI → isAvailable() false and runClaude returns a skipped result; nothing throws.
• bun run typecheck && bun run lint && bun test green; no new runtime dependency.

Stack

All deliverables target the strapped repo (linear stack, each PR based on its parent's branch).

Deliverable	Branch	PR
D1 — eval engine	strapped/evaluation-system/D1-eval-engine	#35
D2 — eval framework	strapped/evaluation-system/D2-eval-framework	#36
D3 — harness eval suite	strapped/evaluation-system/D3-harness-eval-suite	#37
D4 — verified optimizations	strapped/evaluation-system/D4-verified-optimizations	#38